Untitled Document

Computational modeling of neuronal networks is an exciting tool, allowing us to develop and test theories of network function and suggesting novel experiments that we would otherwise not consider. Using morphologically realistic neuron models allows us to be more confident in the accuracy of simulation results; unfortunately, this realism is very computationally expensive. Therefore, we present a general method for reducing morphologically realistic neuron models from hundreds of compartments (full) to 6 (reduced) in three steps, while maintaining the passive, active and synaptic responses of the full model. First, while preserving total surface area, we collapse the proximal and distal halves of the full model’s dendritic tree into 2 and 3 geometrically equal compartments, respectively; the soma remains unchanged, and the axon is ignored. Second, we use an evolutionary algorithm (EA) with a fitness function comparing passive responses of the full and reduced models to find somatic, proximal and distal dendritic values for Rm, Cm, and Ra for the reduced model. Third, we use an EA with a fitness function comparing active and synaptic responses to pick channel conductance densities for the reduced model. We have tested our method on 2 electrophysiologically and morphologically distinct neuron models: a 585 compartment globus pallidus (GP) neuron, and a 517 compartment neuron of the deep cerebellar nuclei (DCN). The average error was less than 5% for both the GP and DCN when comparing the full and reduced somatic spike rate for different somatic (8) and dendritic (4) current injection levels. After aligning spike peaks, the root mean squared error (RMSE) for the spontaneous spike shape was less than 5.6mV for -0.5ms to +1.5ms relative to the spike. With peaks aligned, the RMSE for the afterhyperpolarization trajectory was less than 2.6mV for +1ms to +50ms following the spike. The similarly good results achieved by the application of this method to two very different neuron types attests to the reliability of our generalized approach. Thus, this model reduction method will likely be useful in creating realistic simplified neuron models for network simulations, real-time experiments, and fast searches of parameter space.

Authors:	*E. B. HENDRICKSON1, J. R. EDGERTON2, D. JAEGER2; 1Dept Biomed Engin., Georgia Inst. Technol, Emory Univ., Atlanta, GA; 2Biol., Emory Univ., Atlanta, GA
Computational modeling of neuronal networks is an exciting tool, allowing us to develop and test theories of network function and suggesting novel experiments that we would otherwise not consider. Using morphologically realistic neuron models allows us to be more confident in the accuracy of simulation results; unfortunately, this realism is very computationally expensive. Therefore, we present a general method for reducing morphologically realistic neuron models from hundreds of compartments (full) to 6 (reduced) in three steps, while maintaining the passive, active and synaptic responses of the full model. First, while preserving total surface area, we collapse the proximal and distal halves of the full model’s dendritic tree into 2 and 3 geometrically equal compartments, respectively; the soma remains unchanged, and the axon is ignored. Second, we use an evolutionary algorithm (EA) with a fitness function comparing passive responses of the full and reduced models to find somatic, proximal and distal dendritic values for Rm, Cm, and Ra for the reduced model. Third, we use an EA with a fitness function comparing active and synaptic responses to pick channel conductance densities for the reduced model. We have tested our method on 2 electrophysiologically and morphologically distinct neuron models: a 585 compartment globus pallidus (GP) neuron, and a 517 compartment neuron of the deep cerebellar nuclei (DCN). The average error was less than 5% for both the GP and DCN when comparing the full and reduced somatic spike rate for different somatic (8) and dendritic (4) current injection levels. After aligning spike peaks, the root mean squared error (RMSE) for the spontaneous spike shape was less than 5.6mV for -0.5ms to +1.5ms relative to the spike. With peaks aligned, the RMSE for the afterhyperpolarization trajectory was less than 2.6mV for +1ms to +50ms following the spike. The similarly good results achieved by the application of this method to two very different neuron types attests to the reliability of our generalized approach. Thus, this model reduction method will likely be useful in creating realistic simplified neuron models for network simulations, real-time experiments, and fast searches of parameter space.

Support:	NSF DGE-0333411
	NINDS R01 NS039852